Predicting visit costs with deep neural networks for obstructive sleep apnea patients
School of Business | Master's thesis
Unless otherwise stated, all rights belong to the author. You may download, display and print this publication for Your own personal use. Commercial use is prohibited.
Information and Service Management (ISM)
84 + 1
AbstractIn industrialized countries, sleep apnea affects an estimated 17% of adults. If left untreated, severe sleep apnea can lead to cardiovascular disease or even premature death. In 2019, sleep apnea has attracted a lot of public attention in Finland, in part because Olli Lindholm, a famous Finnish singer, died from it. At the same time, the number of referrals is increasing, putting more and more pressure on healthcare facilities. It makes sense to understand how to allocate resources more efficiently and create budget plans more accurately. Meanwhile, data from electronic health records (EHRs) is growing and machine learning and deep learning data analytics tools are making great strides. There are already some successful applications of deep learning models for predicting treatment outcomes in EHRs. In this study, we aim to train and validate one of the state-of-the-art deep learning models called Transformer to predict visit costs in the next year for patients with obstructive sleep apnea (OSA) using EHRs and compare Transformer with other deep neural networks in model performances. Remote access to the EHRs is provided by Turku University Hospital. First, we perform data preprocessing by selecting data from the EHRs, converting categorical data into ids, and creating input sequences. Then, we create a model architecture with a loss function composed of two sub-loss functions with different scales, including 1) mean squared error (MSE) loss and 2) cross-entropy loss. The MSE loss aims at predicting total costs, while the cross-entropy loss aims at predicting individual costs. As a result, the loss function manages to achieve both objectives with a Top-10 Accuracy of about 94% and an R2 of more than 0.9. Moreover, Transformer outperforms all the other baseline models such as the attention-based encoder-decoder LSTM. Despite the limitations, such as insufficient data, the study shows that EHRs can be one of the sources to support decision making and deep learning models have the potential for resource optimization and budget planning in healthcare.
Thesis advisorSiltala-Li, Lina
Transformer, multiple losses, time series analysis, electronic health records, sleep apnea, machine learning, deep learning, deep neural networks